EEG Strategy Classification

The Problem

When someone multiplies 6 × 4, their brain either remembers the answer or actually computes it. These two strategies look different on an EEG — different timing, different stage patterns. But standard analysis averages hundreds of trials together, which washes out exactly the per-trial variation you'd need to tell them apart.

Nobody had tried using HMP stage patterns as classification features before. So that's what I did.

The Approach

42 participants, 72-channel EEG at 2048 Hz. I ran PCA down to 10 components, split the data participant by participant, then trained a 4-event Hidden Semi-Markov Model only on retrieval trials. The test set was a balanced holdout: held-out retrieval plus procedural trials the model had never seen. If the model really learned what "remembering" looks like, its fit score (Sum Log-Likelihood) should land higher on retrieval than on computation.

The slightly weird part: I repurposed hsmm_mvpy, a package built for cognitive stage recovery, as a classification tool. It wasn't designed for this. That was the whole bet.

The Outcome

AUC = 0.578

Above chance but not useful yet. The retrieval model does assign systematically higher fit scores to retrieval trials — so the signal exists, it's just thin. Single-trial EEG is noisy, SLL wasn't optimized as a classification score, and there's probably genuine overlap between strategies. An honest 0.578 told me more than a cherry-picked accuracy number would have.

Putting It on GitHub

When the project finished I had a pile of scripts, saved intermediates, and a draft report that had drifted from the final public pipeline. I went back through all of it and cleaned it up.

I turned the repo into something a reviewer can parse in one pass: a faster summary, a plain-English method walkthrough, a reproducibility note, and a short explanation of what changed between the draft report and the final public pipeline.

This project is also where I first learned Python, where I worked with genuinely messy real-world data for the first time, and where I had to figure out how to get usable classification output from hsmm_mvpy — a package built for cognitive stage recovery using a cutting-edge HSMM model, not prediction, so getting it to produce numbers I could actually work with took real problem-solving. The result sits somewhere between a proof of concept and an early signal — the dataset wasn't big or clean enough for anything stronger, and SLL as a classification feature has obvious limits. But the method can capture something, and that was the whole point.

What I'd Do Differently